Why this matters
Stripping guardrails from a modern foundation model changes what researchers can probe and what risks an attacker could exploit. This release provides an "obliterated" Gemma 4 E4B variant that intentionally removes refusal behavior while addressing a prior export bug that deleted shared K/V tensors — making the uncensored model usable for local, offline experimentation and adversarial testing.
What Sets It Apart
-
Surgical refusal removal with preservation: the project focuses on removing hard refusal behaviors (reported 0% hard refusals) while ensuring all 720 GGUF tensors remain intact after export. So what: you get an uncensored model that still matches the original tensor layout, avoiding the corruption bug seen in earlier builds.
-
Multiple delivery formats and quant tradeoffs: published as safetensors (bfloat16, ~17 GB) and several GGUF quant variants (Q4_K_M, Q5_K_M, Q8_0, etc.) tuned for llama.cpp/Ollama/phone runtimes. So what: you can run it on desktop inference stacks or on high-end phones with the recommended quant.
-
Practical compatibility notes: Gemma4 is a new architecture; loaders and backends require recent builds (llama.cpp b8665+, updated llama-cpp-python backends, Ollama 0.20+). So what: users must update toolchains to avoid "unsupported architecture" errors and to get stable outputs.
Who It's For and Trade-offs
Great fit if you are doing red‑teaming, safety research, or offline model behavioural analysis and you accept legal and ethical responsibility for uncensored outputs. The release is valuable for testing refusal modalities, prompt‑based attacks, and robustness experiments.
Look elsewhere if you need a production‑safe assistant, content‑moderation guarantees, or a model cleared for user‑facing deployment. Trade‑offs include the underlying 4B model's quality limits (higher soft deflection, occasional repetition), potential for harmful outputs, and legal/ethical liability for misuse. The maintainers explicitly disclaim liability and recommend additional safety layers before any deployment.
Methodological note
The authors report the model was produced using the OBLITERATUS pipeline (aggressive ablation: whitened SVD, attention‑head surgery, winsorized activations) with agentized orchestration. A specific bug was fixed in v3: shared KV weights were projected only once on the owner layer to avoid repeated corruptions that previously caused missing tensors on export. This is a technical release aimed at experimenters, not end users.
